Computation , Constructivism , and Curriculum Design

From the mechanical teaching machines of the early twentieth century to intelligent tutoring systems and the wave of massive open online courses in recent years, many have been motivated by the dream of personalized, adaptive instruction for all students. To achieve this goal, learning scientists and educational technology researchers have largely focused on rule-based systems that rely on extensive domain expertise and constrained machine learning algorithms combined with hand-selected rules for adaptive content selection. While this approach has led to the development of successful intelligent tutoring systems with high quality content, (1) developing such systems can be very costly and (2) they use a very limited form of adaptive content selection. In contrast, some researchers are now starting to apply black box statistical machine learning algorithms in attempting to achieve the dream of adaptive content selection. However, as I will show, these approaches have had relatively limited impact. Instead, I hope to demonstrate that combining insights from both approaches can help in building systems with cost-effective and impactful automated content selection. In particular, I address this using several methods that combine machine learning, human computation, and principles from the learning sciences. First, I will describe how reasoning about model mismatch (i.e., the fact that our statistical models of student learning do not accurately describe how students learn) can help point out limitations in existing approaches and help in creating more robust adaptive content selection policies. Second, I will show experiments that demonstrate how we can use students to create new content in a cost-effective way. In doing so, I will take motivation from the constructivist philosophy of education, whereby I view learner-generated solutions as being a projection of students’ constructions on the written plane, which can then be used to inform other students as they construct their own understandings. Third, I propose to demonstrate how using machine learning can help curate the best learner-generated content. Finally, I propose to show how we can use learning science principles to constrain the search for good activity selection policies. My dissertation aims to demonstrate both how insights from computer science and statistics can inform the learning sciences and how insights from the learning sciences can guide computational approaches with the goal of helping students learn. November 15, 2017 DRAFT